When a pattern is partially anchored and literal prefixes are detected,
we would scan for those prefixes like normal. On a match, we'd verify
whether the text starting at that prefix matched the rest of the regex.
This process doesn't work that well for partially anchored regexes, since
the prefix match presupposes that the regex is allowed to match at that
position. But if, say, a regex like `^z|a` finds a `z` in the middle of
the string, then it has lost the fact that `z` needs to appear at the
beginning of the string, and can therefore falsely report a match.

We could spend some effort and make this case work, but the literal
optimizer is already too complex. We need to simplify it to make future
optimizations like this possible.

Fixes rust-lang#280.